Fast Multiview Co-Clustering in Unified Subspace

Co-clustering has become a research hotspot, which focuses on analyzing block structure and decouple the duality between samples and features, thereby providing a concise approach toward graph-free clustering. Despite this, the label extraction still relies on post-processing, with synergy between independent processes out of evaluation. In addition, while extending it to multiview learning, the redundancy (in high-dimensional feature) and heterogeneity (under different views) of features can lead to difficulty in mining distinct and consensus block structure. In view of these, a novel multiview co-clustering method named Fast Multiview Co-Clustering in Unified Subspace (FOCUS) is put forward, which achieves discrete label decoupling within the same latent space directly. Given that featuring embedding is completed in an unsupervised manner, the principle of information loss minimization is considered to ensure the sparsity and validity of common representations. On this basis, dynamic decoupling is introduced to extract labels for both samples and features, where discrete constraint enables integrated clustering without any post-processing. Besides, extreme feature loss can mislead optimization, so that least-absolute criteria are adopted in function design, while the coupling matrix is further relaxed to be unconstrained for flexible approximation in an enhanced version. In this way, the view weights can be self-updated according to the re-weighted strategy, and the comparison results with eleven state-of-the-art methods on six real-world data sets verify the superiority of our method.